BioML Seminar 5.1 - Danny Reidenbach on Codesigning protein sequence and structure at test-time

BioML Seminar 5.1 - Danny Reidenbach on Codesigning protein sequence and structure at test-time

Title history
  • BioML Seminar 5.1 - Danny Reidenbach on Codesigning protein sequence and structure at test-time — observed on 20261117-0328 UTC (YouTube)
🎙 Danny Reidenbach 👥 14K 📅 November 16, 2026 ⏱ 84 min 👁 56 📄 original study 🧭 2026-11-17
Available in: English (current) Français

Keywords

protein designgenerative modeltest-time scalingwet lab validationlatent space

Summary

The seminar presents Proteina-Complexa, a generative model for protein binder design that codesigns sequence and structure in a continuous latent space, guided by reward-based search at inference time. Danny Reidenbach begins by contextualizing the model within his earlier work (Proteina and Laertina), emphasizing a progression from backbone generation to all-atom monomer design and finally to target-conditioned binder generation. The core innovation is a fused latent representation that integrates discrete residue identity with continuous side-chain structure, enabling joint sequence-structure generation without resorting to inverse folding. The model is trained on a large synthetic dataset generated by cropping domains from AlphaFold database structures, augmenting the limited PDB data and significantly improving performance. At test time, a latent generative search cycle uses physical and structure-based reward models to iteratively refine candidate binders, scaling compute efficiently. The talk includes extensive wet lab validation: over one million designs tested against 127 targets, with binders successfully generated for PDGFR, ActRIIA, kinase targets, a viral glycoprotein, and a free carbohydrate—the first de novo proteins binding a non-protein, polar target. The speaker also discusses architectural choices, such as replacing expensive attention operations with simpler pair-bias transformers, enabling high-throughput screening. Throughout, he answers audience questions about data curation, training lengths, and computational trade-offs. The presentation concludes with a discussion of open challenges and the broader impact of open model families for the scientific community.

225 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides high-value information by presenting a complete pipeline from model design, training, and inference-time search to large-scale experimental validation. The argumentation is solid: the speaker justifies design choices (e.g., continuous vs. discrete latent variables, dropping expensive attention) with empirical evidence and comparisons to prior work like RFdiffusion and BindCraft. The wet lab validation is particularly convincing, as the scale (over a million designs) and novelty (first de novo binders to a free carbohydrate) are clearly articulated. The speaker also addresses potential limitations, such as length constraints and compute costs, showing a balanced perspective.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor through reference to specific prior works (RFdiffusion, BindCraft, AlphaFold) and the speaker’s own published models (Proteina, Laertina), though no external URLs are provided in the description. The validation results are presented with quantitative details, and the speaker is transparent about the messy nature of PDB data and the need for synthetic data augmentation. The title adequately captures the core focus on codesigning sequence and structure with test-time scaling, which is exactly what the talk addresses. The audience Q&A shows engagement and the speaker handles technical questions with expertise, reinforcing the credibility of the content.

209 words

Title / Content Match

The title accurately reflects the content: the talk focuses on the Proteina-Complexa model for codesigning protein sequence and structure with test-time scaling, including wet-lab validation.

Quality & Reliability

8/10

The speaker is a senior research scientist at NVIDIA with a strong academic background, and the talk presents a comprehensive original study including large-scale wet lab validation. The methodology is clear and the claims are grounded in experimental results. However, the presentation is technical and assumes prior knowledge, and some details are simplified.

Key Moments

Contribution & Novelties

The talk presents a novel integration of generative pre-training with test-time scaling for protein binder design, achieving unprecedented scale in wet lab validation. The continuous latent space approach that fuses sequence and structure eliminates the need for inverse folding, simplifying the design loop. The synthetic data augmentation strategy significantly expands the training data beyond the limited PDB. The demonstration of de novo binders to a free carbohydrate is a notable milestone. For those interested in diving deeper, the following resources are recommended:

Pour aller plus loin :

  • AlphaFold — Structure prediction model that underpins many design approaches; its database was used for synthetic data generation.
  • Diffusion model (generative AI) — The generative modeling framework used for protein backbone and latent generation.
  • RFdiffusion — A prior state-of-the-art diffusion model for protein binder design; the talk discusses its limitations and contrasts with Proteina-Complexa.
  • Test-time scaling — The concept from LLMs adapted here to guide protein design through iterative search.

157 words

Radar Profile

The radar profile shows high technical depth and quantity of information, with slightly lower but still strong scores in information quality and reliability. The talk is dense and data-rich, reflecting a expert-level presentation with robust experimental backing.

Reliability 8/10